Skip to content

feat(saturation-v2): add internal EPP queue demand multiplier#1468

Open
ev-shindin wants to merge 1 commit into
llm-d:mainfrom
ev-shindin:feat/epp-queue-multiplier
Open

feat(saturation-v2): add internal EPP queue demand multiplier#1468
ev-shindin wants to merge 1 commit into
llm-d:mainfrom
ev-shindin:feat/epp-queue-multiplier

Conversation

@ev-shindin

Copy link
Copy Markdown
Collaborator

Proposed Changes

  • 🎁 Introduce DefaultEPPQueueDemandMultiplier, an internal (compile-time) knob in the V2 saturation analyzer that scales the token demand estimated from the EPP scheduler queue (requests queued upstream in llm-d flow control) before it is folded into model-level demand.
  • 🧹 Thread the multiplier through estimateSchedulerQueueDemand, applying it to both the input- and output-token components so the model-level total and the per-role (prefill / decode / both) attribution scale consistently.
  • Default value is 1.0, so this is a no-op for current behavior; a follow-up will promote the knob to SaturationScalingConfig for per-model tuning.

Pre-review Checklist

  • E2E tests for any new behavior — N/A: multiplier defaults to 1.0 (no behavior change); covered by a unit test asserting a 2.0 multiplier doubles total and per-role demand.
  • Docs PR for any user-facing impact — N/A: internal knob, not yet surfaced in config.
  • Proposal PR for any new enhancement or change to existing behavior — N/A: no observable behavior change at the default; config-facing proposal will accompany the follow-up.

Introduce DefaultEPPQueueDemandMultiplier, an internal knob that scales
the token demand estimated from the EPP scheduler queue (requests queued
upstream in llm-d flow control) before it is folded into model-level
demand. The multiplier is applied to both input and output token
components in estimateSchedulerQueueDemand so the model-level total and
per-role (prefill/decode/both) attribution scale consistently.

Default is 1.0 (no behavior change). This is a compile-time constant for
now; a follow-up will promote it to SaturationScalingConfig for per-model
tuning.

Signed-off-by: Evgeny Shindin <evgensh@il.ibm.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant